The Nature of the Next Step: An EMG-based Machine Learning Approach To Classifying Locomotion Modes During Continuous Walking

Lower limb loss is prevalent in the USA and world at large. Despite this, there are currently no commercially available bionic lower limb prosthesis which implement some kind of neuromuscular activity in their control. There are significant differences between different types of steps with regards to several factors such as ground reaction forces, stability and stance time length. As poor mobility is one of the leading reasons for lower limb prosthesis device abandonment, it is beneficial to develop a bionic lower limb prosthesis capable of providing suitable assistance depending on the nature of the step (backwards, forward, stair climbing, turns) to be completed. However, in order to achieve this there must first exist a reliable way to classify the nature of the step to be completed. Thus, this research will focus on the development of an EMG-based machine learning algorithm capable of identifying the nature of the next step during continuous walking. Surface Electromyography (sEMG) signals will be collected from four participants without disabilities that might affect gait. A weighted consensus model consisting of supervised machine learning algorithms will be constructed to identify the nature of the step using sEMG data from -40% to 80% of the gait cycle. Two methods for constructing the weighted consensus model will be considered – the first, a model that uses the top performing models of a specific subject in the final model for that subject and the second, a model which uses the best performing models on average. Two methods will also be considered for training the models, one where training is done individually per subject and one where all training set data is combined and concurrently used for training. Upon the conclusion of this research, it will be determined that the weighted consensus model identified is only capable of accurately classifying the step nature into stair descent, backwards waling and a third category consisting of turns, stair ascent and forward level walking. Future work should focus on the development of a cascading model which first applies the weighted consensus model identified in this research and then attempts to reclassify the steps belonging to the third category into their respective classes.